UC San Diego

Cardiac MRI is the gold standard for quantification of cardiac volumetry, function, and blood flow. Despite the wealth of information that may be gleamed from these acquisitions, its use has been limited primarily to academic and specialty clinics due to the need for specialty trained physicians and technologists required for planning of these scans.Recently, deep convolutional neural networks (DCNNs) have shown promise in automating various aspects of radiological workflows, such as landmark localization. However, a primary limitation to applying DCNNs to clinical practice include uncertainty of how well an algorithm will perform outside of the environment in which it was trained. Moreover, these systems are often seen as “black boxes”, which fail to provide an explanation of how an answer was achieved. Providing a way in which clinical end users may have confidence in these systems is therefore essential for clinical adoption of any medically focused DCNN system.

With these concerns in mind, I explore the potential automating the planning of Cardiac MR imaging planes using DCNN. In the first chapter, I explore the potential of automating the prescription of long-axis and short axis imaging planes by localizing the landmarks. To preserve the iterability in the DCNN, I regress pseudoprobability heatmaps (termed heatmap regression) centered at the valve and apex landmarks. I demonstrate that this approach of heatmap regression not only accurately identifies the landmarks, it is additionally able to recreate imaging planes similar to those defined by the ground truth landmarks or those acquired by a technologist at the time of original acquisition.

In my second chapter, I explore the potential to applying these DCNNs within a clinical setting. I first established the importance of our angulation metric for assessing the accuracy imaging plane. To assess the generalizability of this system to different clinical environments, I calculated the angulation error between ground truth defined and DCNN predicted imaging planes. As an additional level of comparison, angulation error was calculated for technologist acquired imaging planes. I found that this system of DCNNs generally achieved similar or better performance compared to a technologist.

Finally, in my third chapter, I explore the potential of adapting my DCNN algorithm to different clinical environments, using the differences in imaging characteristics seen at 1.5T vs 3T as a model system. To achieve this, I developed a methodology for selecting cases with greatest model uncertainty for transfer learning. I moreover developed two novel uncertainty metrics based either strength of prediction or test-time augmentation spatial variance pseudoprobability maps. To assess the performance of this approach, I used a model trained on only 1.5T long-axis images, and calculated pseudoprobability metrics of 3T long-axis images. We assessed the potential of each pseudoprobability metric by ranking 3T long-axis images by either increasing, decreasing, or random values. I found that 3T images with the highest uncertainty most efficiently increased the transfer learning data-efficiency for the apex, consistent with a good uncertainty metric. Moreover, I found that incorporation of 1.5T data into the transfer learning process helped preserve the initial performance at 1.5T.